Research Foundation

Aureonics Research

The mathematical framework behind constitutional AI governance. Formalized, falsifiable, and openly reported — every result stated exactly as far as it is proven, and no further.

📄

Aureonics: Constitutional Triadic Framework for Stable Adaptive Intelligence

Emmanuel King · Independent Research · Lagos, Nigeria

AuthorEmmanuel King · Lagos, Nigeria
AccessOpen-access preprint · Zenodo
Read Full Paper ↗
Abstract

We present Aureonics, a constitutional triadic framework for stable adaptive intelligence. The framework models an AI system’s constitutional health as a point on the probability simplex over three irreducible invariants — Continuity (C), Reciprocity (R), and Sovereignty (S) — constrained so that C + R + S = 1. The stability margin M = min(C, R, S) is a scalar measure of constitutional health. A Control-Barrier-Function governor detects constitutional drift before failure and applies mass-conserving corrections that keep the state in a safe interior set. The framework is operationalized as the PRAXIS pipeline with cryptographic audit receipts, a z-weighted Lyapunov certificate, and per-turn constitutional measurement. The system is mathematically bounded and falsifiable; below we state precisely which properties are proven, which are numerically certified, and which remain open.

Mathematical Framework

The state, the governor, the certificate

x = (C, R, S), C + R + S = 1

Constitutional state — a point on the 2-simplex. Every governor operation preserves the sum-to-one constraint exactly.

M(x) = min(C, R, S)

Stability margin — the system is only as stable as its weakest constitutional pillar.

Gᵢ(x,T) = kᵢ · (φᵢ − φ̄)

Governor force on pillar i — a mass-conserving push (Σ Gᵢ = 0) toward the safe interior, proportional to how far that pillar sits below the target.

kᵢ(x,T) = k₀ · wᵢ(T) / (M(x) + εₖ)

Adaptive stiffness — correction strength grows as the margin M shrinks, so the governor pushes hardest exactly when the state is closest to the boundary.

V_z(x) = −Σ zᵢ·log(xᵢ) + (μ/2)·Σ max(0, τ−xᵢ)²

The z-weighted log-barrier Lyapunov certificate (§11). An interior-point barrier plus a quadratic penalty active only inside the safety margin τ — the same structural family as a control barrier function (Ames et al., 2019). Stamped on every governed receipt.

ẋ = −Π_Σ ∇V_z(x) ⟹ V̇_z ≤ 0

Under the idealized continuous flow (projected gradient descent of V_z onto the simplex), the certificate is non-increasing — a standard Lyapunov descent argument. This is the proven single-pillar result.

Π_S(x) = argmin_{y∈S} ‖y − x‖₂, S = {y : Σy=1, yᵢ ≥ τ}

Exact Euclidean projection onto the floor-constrained simplex (Duchi–Shalev-Shwartz–Singer). Non-expansive, and it makes forward invariance of the τ floor hold by construction — the deployed governor and the offline simulator both use it.

z_{t+1} = normalize(clamp(ρ·z_t + (1−ρ)·x_t − A(t), τ/2, 1−τ))

Constitutional memory — the z-weight update. A Banach contraction (ρ = 0.85): bounded (Theorem 3a) and convergent (Theorem 3b). Lets the governor respond to sustained pressure on a pillar, not just the current turn.

Constitutional Constants

Frozen parameters

These values are fixed in code and never tuned per request. The health bands below are defined entirely by them.

τ_floor0.05
CBF hard floor
τ_LYP0.08
Lyapunov-penalty band onset
τ_recovery0.15
recovery floor
k₀0.30
base governor stiffness
εₖ0.01
stiffness regularizer
ρ0.85
z-update contraction rate
Health bands: OPTIMAL M ≥ 0.25 · ALERT 0.15 ≤ M < 0.25 · STRESSED 0.08 ≤ M < 0.15 · CRITICAL M < 0.08
Formal Stability

What is proven, certified, and open

Constitutional state is a point on the simplex; safety is enforced by a barrier function; stability is argued with a Lyapunov function. We are precise about the strength of each claim.

Single-pillar Lyapunov descentProven

Under the continuous flow ẋ = −Π_Σ∇V_z, V̇_z ≤ 0 in the single-pillar violation regime (condition k₀/εₖ > 3B/2 satisfied with ~20× margin). This is the result external-facing copy refers to when it says 'Lyapunov stability'.

Forward invariance of the τ floorBy construction

The exact floor-constrained simplex projection returns xᵢ ≥ τ for every pillar, every step — so the governed state provably never leaves the safe set. Verified in the live governor and the offline simulator alike.

CBF simulator — numerical FPL-1 certificateCertified (numerical)

The offline governed-vs-ungoverned simulator (below) certifies its governed arm as LYAPUNOV STABLE + FORWARD INVARIANT at the continuous-flow limit: Lyapunov descent ratio > 0.6, zero floor incursions, bounded V_z excursion < 0.25, across seeds. This is a seeded, finite-horizon NUMERICAL certificate — strong evidence for the flow, not a replacement for the analytical proof below.

Multi-pillar global Lyapunov proofAdvanced · open

Substantially advanced 2026-07-21, not closed. (a) The idealized flow ẋ=−Π∇V_z is globally, multi-pillar Lyapunov-stable because V_z is convex on the floor-simplex (V̇_z=−‖Π∇V_z‖²≤0 to a unique minimizer). (b) The deployed governor's action on V_z is proven non-positive for all states including two-pillars-stressed, via Chebyshev's sum inequality — so multi-pillar is not a new structural obstruction. The sole residual is the quantitative governor-vs-drift margin, the same condition already discharged single-pillar. The analytical multi-pillar theorem is not complete, and we never claim otherwise.

Live Counterfactual

Governed vs ungoverned — the thing production can't show

Production only ever runs with the barrier active, so a real user can never be shown what happens without it. This controlled simulation runs the identical perturbation sequence twice from one seed — once governed, once not — and certifies the governed arm’s stability. It updates live from /api/cbf-simulation.

Initializing Kernel Simulation...

Reproduce the classification and the discretization analysis behind it: npx tsx scripts/cbf/fpl1-dt-sweep.ts. Full method and results in research/empirical-results.md (Run 002).

Open Problems

Stated honestly — science that can be falsified

Problem 1 — Global Lyapunov proofPartial

Single-pillar regime proven; multi-pillar simultaneous violation open. Approach: comparison system or LaSalle invariance, leveraging non-expansivity of the Duchi projection. Priority: medium.

Problem 2 — Nonlinear Pareto frontierOpen

Full characterization under λ > 0, including the coupling to the adaptive floor τ_eff(z, ℓ). Approach: Lagrangian methods + numerical continuation.

Problem 3 — Complete z-update rulePartial

State-space side closed (velocity, n_stable, drift, σ_viol, attack_pressure specified and proven bounded/convergent). Remaining: characterize the dp_attack/dt coupling to law events as a hybrid dynamical system over three margin regions.

Falsifiable Predictions

P1–P12

A framework that cannot be falsified is not science. Twelve pre-registered predictions; their status is reported exactly, including “untested.”

P1–P9 — original paper predictions. Untested at scale.

P10 — per-session adversarial collapse: repeated adversarial turns within one session collapse M faster than the same prompts across independent sessions (attack_pressure raises the effective floor over time). Proposed, testable on SSS50.

P11 — faster slow-drip detection: time-to-detection is shorter when σ_viol accumulates at τ_LYP (0.08) than at τ_floor (0.05). Proposed.

P12 — taxonomy partition completeness: the empirical distribution of fired laws on production traffic matches the assumed attack partition, with no residual “other” class. Proposed.

Empirical Evidence

Adversarial benchmarks

The governor is developed against a 550-vector internal adversarial suite (8 attack classes) and evaluated on external public benchmarks under symmetric judging — the bare and governed arms scored by the same external judge on their actual output text.

920+
external adversarial prompts (HarmBench / JailbreakBench / AdvBench)
0%
governed attack-success rate reported on those runs
550
internal adversarial vectors, 8 attack classes
37,701
production turns with logged ΔV_z sign

Honest status: automated benchmark re-runs are currently paused to control free-tier provider quota; the figures above are from published runs, and the harness is fully reproducible (see the benchmarks page for live, dated results as they publish). Attack-success is measured over harmful prompts only; over-refusal on benign prompts is reported separately, never netted against it.

Governance Pipeline

PRAXIS — every prompt, every time

01Pre-eval classification (CLEAR / HIGH)
02Embedding + constitutional memory recall
03z-trajectory update (proven Banach rule)
04SovereignKernel — compute C, R, S; enforce C+R+S=1
05Dual inference — bare vs governed, same model
06Governor fires if M < τ — project to safe interior
07SHA-256 audit receipt (input, output, bound state)
08Constitutional output + cryptographic proof
Attack Taxonomy
bypass_attempt
S collapse
identity_reframe
C collapse
sycophancy
R collapse
multi_attack
all pillars
slow_drip
M global
Reproducibility

Verify it yourself

Every governed turn writes an append-only SHA-256 receipt binding the input hash, output hash, and constitutional state — independently re-verifiable at /api/lex/verify.

The stability certificate is a pure, seeded function — reproduce it with npx tsx scripts/cbf/fpl1-dt-sweep.ts.

The full method and per-run analysis live in research/empirical-results.md and research/open-problems.md.

Cite This Work
@misc{king2026aureonics,
  title  = {Aureonics: A Constitutional Triadic Framework
            for Stable Adaptive Intelligence},
  author = {Emmanuel King},
  year   = {2026},
  doi    = {10.5281/zenodo.18944242},
  url    = {https://doi.org/10.5281/zenodo.18944242},
  note   = {Independent researcher, Nigeria.
            ORCID: 0009-0000-2986-4935}
}